Deep Learning Modeling for Prediction of Cognitive Task Related Features from Resting-state fMRI Data

Abstract

In recent years, several neural network methods, such as brain functional networks, have been proposed to efficiently learn non-Euclidean graph structures. In this study, we modified the connectivity-based graph convolutional network (cGCN) developed by Wang et al. (2021) for autism spectrum disorder (ASD) classification into a regression model and used resting-state functional magnetic resonance imaging (rs-fMRI) data to predict the scores on the offline Kohs block-design test for a total of 615 subjects aged 33–89 years. To convert from discrimination to regression, we employed a technique that introduces a fully connected layer in the cGCN and connects the long short-term memory (LSTM) in the last output phase instead of the Softmax layer. The results showed that our cGCN–LSTM was more accurate than the baseline LASSO regression model, and that the predictions correlated significantly with the measured scores of the cognitive function test. Moreover, we used the leave-one-out and leave-two-out occlusion methods to extract important regions of interest (ROIs), as well as networks from the model. It was acknowledged that the Kohs block-design test scores were negatively correlated with age, but the results suggested the possibility of age-related cognitive decline that could not be captured by age prediction models alone. We found that only the nodes of the default mode network and cerebellum contain some significant within-networks; however, overall, between-network connectivity overwhelmingly contributed to the prediction regardless of the weight of the role in the age projection. This model and the leave-two-out occlusion method allowed us to identify the regions and networks involved in further task-based fMRI experiments in advance. Our methodology has the potential to make the design of task fMRI experiments more rational and accurate before planning and conducting actual scans.